Evidence-Type Competition: When Can Interventional Data Teach Language Models Causal Direction?

За 30 секунд

Что произошло

Research shows language models trained on interventional causal data still copy observational correlations at inference time, contradicting assumptions about causal learning.

Почему это важно

Matters for engineers building systems requiring reliable causal reasoning, especially when training data mixes observational and interventional examples.

На что обратить внимание

The findings use synthetic environments with Simpson's paradox; real-world applicability and scaling behavior beyond 0.93B parameters remain unclear.